Creating and Exploring a Large Photorealistic Virtual Space
Josef Sivic1 ∗, Biliana Kaneva2 , Antonio Torralba2 , Shai Avidan3 and William T. Freeman2
1
INRIA, WILLOW Project, Laboratoire d’Informatique de l’Ecole Normale Superieure, Paris, France
2
CSAIL, Massachusetts Institute of Technology, Cambridge, MA 02139, U.S.A.
3
Adobe Research, Newton, MA, U.S.A.
The supplementary video can be viewed at:
http://people.csail.mit.edu/josef/iv08/video.avi
Abstract
showing the retrieved images as thumbnails, one can display
images in a 3D context, enhancing the user experience. Instead of typing query tags, the user can use simple and intuitive 3D controls to navigate. Indeed, the Street View approach offered by several companies allows users to virtually
wander the streets of a city in a 3D like fashion. This was underscored by the success of the PhotoTourism interface [27]
which put all images of a given scene in a common 3D space
using bundle adjustment. Users browse the image collection
by moving in 3D space.
If all images are collected from the same point of view
then a big Gigapixel image can be constructed and viewed
interactively using familiar 2D image controls [16]. Alternatively, if the image dataset consists of a random collection of
images of different scenes, taken at different times, one can
construct an AutoCollage [25]. This gives visually pleasing
collages, but fails to scale to large collections where thousands or millions of images are involved.
Another option considered in the past is to treat images
as points in a high dimensional space, compute the distance
between them and use multi-dimensional scaling to display
them in the image plane for the user to navigate through [26].
However, there is no effort to create a virtual 3D world and
as a result there is no sense of ”being there”.
In this work, we use intuitive 3D navigation controls to
explore a virtual world created from a large collection of images, as illustrated in figure 1. This side-steps the problem
of image tagging altogether, relying instead on an automatic
image similarity measure. In our approach the user is free to
move from one image to the next using intuitive 3D controls
such as move left/right, zoom and rotate. In response to user
commands, the system retrieves the most visually similar images, from a data set of several million images, and displays
them in correct geometric and photometric alignment with
respect to the current photo. This gives the illusion of moving in some large virtual space. We preprocess the collection
ahead of time, allowing the system to respond interactively
to user input.
There are a number of applications to this technology.
We present a system for exploring large collections of
photos in a virtual 3D space. Our system does not assume
the photographs are of a single real 3D location, nor that
they were taken at the same time. Instead, we organize the
photos in themes, such as city streets or skylines, and let
users navigate within each theme using intuitive 3D controls
that include move left/right, zoom and rotate. Themes allow us to maintain a coherent semantic meaning of the tour,
while visual similarity allows us to create a ”being there”
impression, as if the images were of a particular location.
We present results on a collection of several million images
downloaded from Flickr and broken into themes that consist
of a few hundred thousand images each. A byproduct of our
system is the ability to construct extremely long panoramas,
as well as image taxi, a program that generates a virtual
tour between a user supplied start and finish images. The
system, and its underlying technology can be used in a variety of applications such as games, movies and online virtual
3D spaces like Second Life.
1. Introduction
Exploring large collections of images requires one to determine how to organize images and how to present them.
A leading approach to organizing photo collections is to use
tags. The user navigates the collection by typing a query
tag and reviewing the retrieved images. Typically, the results
are shown as a page of thumbnails or a slide show, which is
a practical but not engaging way to browse images. Moreover, automatic image tagging is still an unsolved problem
and manual image tagging is a time consuming process that
does not always occur in practice.
Of particular interest are geo-referenced images that can
be automatically and accurately tagged with GPS data. This
allows users to explore the collection based on location, but
more importantly offers a new way to navigate. Instead of
∗ This
work was carried out whilst J.S. was at MIT, CSAIL
1
Input Image
Move Left
Move Left
Input Image
Move Forward
Move Forward
Figure 1: Given the user supplied starting image, our system lets the user navigate through a collection of images as if in a 3D world.
First, imagine large 3D virtual online worlds like Second
Life, where users are free to roam around. We make it possible for them to tour a photorealistic environment, instead
of a computer generated one. Alternatively our method can
be used to construct large photorealistic environments for
games. Other applications include movies or advertisements,
where very wide backgrounds for a particular scene may be
needed.
A byproduct of our method is the ability to create different types of panoramas such as infinitely long panoramas
or an image taxi. Image taxi allows the user to specify the
first and last images of a tour and the program will automatically generate a virtual tour through the image collection that
starts and ends at the user specified images. This extends the
large body of work on multi-perspective panoramas [22, 24]
where a large number of images taken from different view
points are stitched together. Our method does not assume
images to be of the same scene or taken at the same time.
In a similar fashion we can create infinite zoom effects
that resemble the ”Droste effect”1 . The effect is named after
a particular image that appeared on the tins and boxes of the
Dutch Droste cocoa powder. It displays a nurse carrying a
serving tray with a cup of hot chocolate and a box of Droste
cocoa. The recursive effect first appeared in 1904, and eventually become a household notion. This effect has been used
by various artistic groups to generate infinite zooms2 .
To achieve these goals we use a number of components.
First, we use a large collection of images, which defines the
image space the user can explore. Second, we use a convenient interface to break images into themes to ensure the
semantic coherence of the tour. Third, we use an image similarity metric that allows us to retrieve images similar to a
transformed version of the current image (e.g. to support a
”move left/right” operation we need to shift the current image, before retrieving similar images). Finally, we show how
to combine two images in a visually pleasing manner to create the illusion of moving in a single, coherent 3D space.
2. Constructing the image space
In this section we describe how we collect and arrange a
large collection of images into a virtual 3D space for the user
to explore as if walking through the pictures. The images do
not need to correspond to a real unique 3D space as in Phototourism [27]. Instead, our images are expected to correspond
1 http://en.wikipedia.org/wiki/Droste
effect
2 http://zoomquilt2.madmindworx.com/zoomquilt2.swf
to unrelated places.
We collected a dataset of more than 6 million images
from Flickr by querying for suitable image tags and relevant
groups. We queried for particular locations in conjunction
with tags such as ’New York street’, or ‘California beach’,
and also downloaded images from Flickr groups returned by
search on terms relevant to particular themes, such as ‘Landscape’ or ‘Forest’. The typical resolution of the downloaded
images is 500 × 375 pixels. One million jpeg compressed
images takes about 120GB of hard-disk space.
Traditional image retrieval deals with the task of finding
images that are similar to a particular query image. Here we
want to extend this query by also allowing transformations
between images. For instance, we want to find images similar to a photo taken if we rotate the camera by 45 degrees
with respect to a query image. We hope that the candidate
set contains images that can be seamlessly blended with the
query image after applying the appropriate camera transformation. We show how a simple planar image model can be
used to retrieve images that are related to the query image by
3D transformations.
2.1. Geometric scene matching
Our goal is to extend a given image using images from the
theme database to give an impression of a particular camera
motion. We consider three camera motions: (i) translation
left/right, (ii) horizontal rotation (pan), and (iii) zoom (forward motion). The camera motions are illustrated in figure 2.
First, we synthesize a new view of the current image as seen
from the new desired camera location. Camera translation
is approximated by a translation in the image plane, ignoring parallax effects, horizontal camera rotation is achieved
by applying appropriate homography transformation to the
image and zoom is approximated by scaling the image.
Now, we find semantically similar images in the database
coarsely matching the geometry of the observed portion of
the synthesized view. For example, when the camera rotation changes a fronto-parallel view of a building to a view
with a strong perspective (as shown in the middle column of
figure 2), we find most retrieved images depict scenes looking down a street.
2.2. Image representation
A key component of our approach is finding a set of semantically similar images to a given query image. For example, if the query image contains a cityscape in a sunset with a
park in the foreground, we would like to find a candidate set
Input image
Input image
Input image
Translation
Camera
rotation
Best match translation
Best match rotation
Stitched translation
Forward motion
Best match forward
Stitched rotation
Stitched zoom
Figure 2: Scene matching with camera view transformations. First
row: The input image with the desired camera view overlaid in
green. Second row: The synthesized view from the new camera.
The goal is to find images which can fill-in the unseen portion of
the image (shown in black) while matching the visible portion. The
third row shows the top matching image found in the dataset of
street scenes for each motion. The fourth row illustrates the induced
camera motion between the two pictures. The final row shows the
composite after Poisson blending.
of images with similar objects, camera viewpoint and lighting. We describe the image features we use to achieve this
and how to extract semantic scene description and camera
properties from them.
Semantic scene matching is a difficult task but some success has been recently shown using large databases of millions of images [13, 29]. In this paper we show that we
can also induce camera transformations without an explicit
model of the 3D geometry of the scene. Matching results
are sometimes noisy; for example, a river is sometimes mismatched for a road. In a recent work on image completion
[13] this issue was addressed by relying on user interaction,
essentially allowing the user to select a visually pleasing result among a set of candidate completions. We wish to find
matching images automatically or with minimal user interaction. To reduce the difficulty of scene matching we train
classifiers to pre-select images of particular scene types or
themes from the entire image collection. The user can then
navigate in a visual space that is built only from images of
a particular theme or a combination of themes. Images are
represented using the GIST descriptor, which was found to
perform well for scene classification [21]. The GIST descriptor measures the oriented edge energy at multiple scales
aggregated into coarse spatial bins. In this work we use three
scales with (from coarse to fine) 8, 8 and 4 filter orientations
aggregated into 6 × 4 spatial bins, resulting in a 480 dimensional image descriptor. In addition we represent a rough
spatial layout of colors by down-sampling each of the RGB
channels to 8 × 8 pixels. We normalize both GIST and the
color layout vectors to have unit norm. This makes both
measures comparable.
As illustrated in figure 2, not all pixels are observed in
the transformed image and hence only descriptor values extracted from the observed descriptor cells in the query image
are considered for matching. In this case, the GIST and the
color layout vectors are renormalized to have unit norm over
the visible region.
The distance between two images is evaluated as the sum
of square differences between the GIST descriptors and the
low-resolution color layout of the two images. We set the
weight between GIST and color to 0.5, which we found is a
good compromise between matching on the geometric structure of the scene captured by GIST and the layout of colors.
Currently we consider only images in the landscape format
with an aspect ratio close to 4:3. This represents about 75%
of all collected images.
Figure 3 shows an example of a query image, the bins
used to compute the image descriptor and the closest matching images from a dataset of 10,000 street images. The images returned belong to similar scene categories and have
similar camera properties (same perspective pattern and similar depth).
2.3. Alignment and compositing
Images returned by the scene matcher tend to contain similar scenes, but can still be misaligned as the GIST descriptors matches only the rough spatial structure given by the
6 × 4 grid of cells. For example, in the case of city skylines,
the horizon line can be at a slightly different height.
To fine tune the alignment, we apply a standard gradient
descent alignment [20, 28] minimizing the mean of the sum
of square pixel color differences in the overlap region between the two images. To do this, we search over three parameters: translation offset in both the x and y direction and
scale. The alignment is performed on images down-sampled
to 1/6 of their original resolution. In the case of translations
and zoom, the alignment search is initialized by the synthesized view transformation. In the case of camera rotation we
initialize the alignment with a translation in the x direction
matching the image overlap induced by the rotation, e.g. half
the image width for the example in middle column of figure 2. As a result, the camera rotation is approximated by a
translation and scale in the image domain. The camera rotation is only used in the scene matching to induce a change in
the geometry of the scene.
The aligned images are blended along a seam in their re-
Query image
Query image
GIST
Camera rotation & GIST
Best match
Top matches
Top matches
Best match after rotation
Figure 3: Each row shows the input image, the 6 × 4 spatial bins used to compute the GIST description, the best match on a dataset of 10,000
images, and the next 6 best matches. Top row: we look for images that match the full GIST descriptor. Bottom row: Result of a query after
simulating a camera rotation. The returned images contain a new perspective, similar to the one that we will have obtained by rotating the
camera 45 degrees to the right.
gion of overlap using Poisson blending [23]. In the case
of camera translation and rotation we find a seam running
from the top to the bottom of the overlap region minimizing
the sum of absolute pixel color differences by solving a dynamic program [10]. In the case of zoom, where images are
within each other, we find four seams, one for each side of
the overlap region. In addition, to preserve a significant portion of the zoomed-in image, we constrain each seam to lie
close to the boundary of the overlap region. Finally, images
are blended in the gradient domain using the Poisson solver
of [2].
To obtain a sequence of images with a particular camera motion the above process is applied iteratively using the
most recently added image as a query.
2.4. Estimation of camera parameters
Although we do not perform a full estimation of the 3D
geometry of the scene structure, we show that the estimation
of some simple camera parameters can improve the quality
of results when navigating through the collection. Specifically, we show how to estimate 3D camera properties from
a single image using the GIST descriptor. Here we follow
[18, 30] in which the estimation of the horizon line is posed
as a regression problem. As opposed to [7, 9] in which camera orientation is estimated by an explicit model of the scene
geometry, we use machine learning to train a regressor. The
main advantage of our method is that it works even when the
scene lacks clear 3D features such as a vanishing point. We
collected 3, 000 training images for which we entered the location of the horizon line manually (for pictures taken by a
person standing on the ground, the horizon line can be approximated by the average vertical location of all the faces
present in the image). We use weighted mixture of linear
regressors [12] to estimate the location of the horizon line
Forward motion
Input image
Forward motion on ground plane
Input image
Best match forward
Query region
Best match forward
Query region
Figure 4: The top row shows the queried region when we take the
central image portion. The bottom row shows the results obtained
when centering the queried region on the horizon line. The retrieved
images contain roads with similar perspective to the input image.
from the GIST features as described in [30].
Figure 4 illustrates the importance of estimating the location of the horizon line. In this example, we want forward
motion to represent a person walking on the ground plane.
As shown in the top row, if we zoom into the picture by using the image center as the focus of expansion, we move into
the sky region. However, if we zoom in on the horizon line
we simulate a person moving on the ground plane. In order
to generate a motion that simulates a person moving on the
ground plane it is important to keep the horizon line within
the queried region.
2.5. Organizing pictures into themes
When performing a query after a camera motion, the information available for matching is weaker than the original GIST descriptor (due to the smaller image overlap after
the transformation). This results in semantically incoherent
Skyline
Graph
Landscapes
People
Streets
Figure 5: Example of images belonging to different scene themes.
Partitioning a large collection of images improves the quality of the
results. The classification is done automatically. When navigating
through the image collection, it is important to keep the themes constant when moving from one picture to another to avoid undesired
transitions. The user can also allow transitions across themes.
transitions between images when creating long sequences (in
fact, out of 10 sequences none of them produced satisfactory
results when using a random set of 0.5 million images). In
order to enhance the GIST descriptor, we use the theme as
an additional constraint. The theme of an image is, in most
cases, invariant with respect to camera motions. Therefore,
we can use this as an additional constraint by only matching
pictures that belong to the same theme. Using the themes
results in dramatically improved results. It also lowers the
memory and CPU requirements as only part of the database
needs to be searched. Examples of themes we consider in
this work are shown in figure 5.
To obtain a visually and semantically coherent set of images for each theme we train theme classifiers from manually labeled training data in a manner similar to [5, 19].
For each theme we train 1-vs-all linear SVM classifier from
about 1,000 positive and 2,000 negative training images. We
have developed a suitable labeling interface so that the classifier can be trained interactively. At each iteration the user
can visually assess the classifier performance and label additional images if needed. At each iteration, the most uncertain
images are shown to the user for labeling.
3. Navigating the virtual 3D space
We process the image database and create a graph where
the nodes represent images and there are different types of
(directed) edges, corresponding to different types of motion,
connecting the nodes. The weight of each edge corresponds
to the matching cost between the two images and the particular camera motion associated with this edge. We typically
retain only the top 10 matches for every particular motion
(Figure 6). This gives us a sparse graph that allows our system to tour the virtual space efficiently. The main bottleneck
Virtual 3D world
Figure 6: In the original graph, each image is a node and there are
different types of edges (color coded in the upper left graph) that
correspond to different camera motions. For each motion, there
are many good matches and we typically keep just the top 10 for
each edge type. All the images of a collection can by organized in a
virtual world by greedily assigning to each image the best neighbors
and camera motions that minimize the matching cost.
in processing the database is to compute the GIST descriptors for every image, which we do at about 5 images per
second. Once the image database is processed and the graph
is constructed the user can start navigating in it using intuitive 3D controls. There are different modes of operation, all
relying on the same basic machinery. Given a query image,
the user can interactively tour the virtual space, ask for a predefined path or generate a tour between two images using the
image taxi metaphor. It takes about a second to retrieve the
matching images, in our matlab implementation, and a couple of seconds to align and blend the retrieved images with
the query image.
3.1. Sequences with different camera transformations
The user selects an image that serves as his gateway to the
virtual 3D space and starts navigating interactively in a particular theme using intuitive 3D commands (move left/right,
zoom, rotate). The system currently can either choose the top
matching image automatically or let the user choose the best
one from the top 5 matches that were found. The tour can
be recorded and then shown as a video, seamlessly moving
between different images.
In some cases, users are interested in a particular type of
tour. For example, the user might be interested in generating
a tour that starts with a short forward motion, followed by
a long left translation and concluding with a right rotation.
The system will then query the graph and construct a video
sequence that satisfies the user’s request. In addition, the
user can choose the image taxi option. In this case, the user
only specifies the first and last images of the tour and the
image taxi takes the user along the shortest path in the graph
connecting these two images.
Figures 7, 8(a) and 8(b) show image sequences for camera zoom, translation and rotation respectively obtained from
a ‘street’ theme. Note that the left/right translation motion tends to preserve the camera orientation with respect to
the scene (the scene remains roughly fronto-parallel), while
the rotation motion induces a change in perspective between
consecutive images (Fig 9). The translation and rotation sequence were produced fully automatically. The zoom-in sequence was produced interactively by letting the user specify
the zoom-in direction - toward the horizon of the image.
(a)
(b)
(c)
(d)
The reader is encouraged to view the accompanied video
that presents the different tours generated by our system. All
the sequences in the video were generated automatically, except for the “Hollywood” and “Windows” sequence that had
the user choose the best match out of the top 5 candidates.
To generate the motion videos we produce the composite
(including Poisson blending) of each image of the sequence
with its neighboring images. The intermediate frames of the
motion are synthesized from these composited images by
temporal blending to hide small color differences resulting
from Poisson blending subsets of images independently. In
the case of translation and rotation we look at only one image to each side of the current image. In the case of zoom,
a particular composited image can contain pixels from several images ahead, as illustrated in figure 7(d). The translation is simulated by translating with constant per pixel speed
between consecutive images of the sequence (also applying
small scale changes if the consecutive images are of different size). The case of rotation is similar to translation but, in
addition, we display the resulting images on a cylinder. In
the case of zoom we apply constant scale change between
consecutive frames of the video to create an impression of
a person walking at a constant speed in the 3D world. The
zoom is performed toward the center of the next image in
the sequence, which can result in (typically small) changes
in zoom direction between consecutive images.
Infinite panoramas and the infinite street: As a byproduct of our system we can construct infinite panoramas by
automatically moving the camera at fixed intervals. Fig. 7
gives an example of an infinite perspective street generated
with our system. Figure 10 shows examples of long panoramas created for various themes.
The image taxi: finding a path between two images:
The image taxi allows the user to specify start and end images and let the system take him on a tour along the shortest
path between them. Given two input images, we first connect
them to the graph and then find the shortest path between
them using the Dijkstra algorithm [6]. We then follow the
path creating a tour based on the different edges along the
way.
(e)
Figure 7: Zoom sequence of 16 images. (a) The query image. (b)
The query image composited with the consecutive images in the
sequence. (c) Image boundaries of the following images in the sequence overlaid over the composited image (b). (d) Masks indicating areas in (b) coming from different images. (e) Images 2–4
used to generate the zoom sequence. Each consecutive image was
obtained by inducing camera zoom as illustrated in figure 2. The
zoom direction was indicated interactively by clicking on the horizon line.
Figure 9: A camera rotation can be used to generate from a single
picture a good guess of the surrounding environment not covered by
the camera. Here, the system is hallucinating what could be behind
of the camera (original image marked with a frustrom). Note that
the back of the camera is also a perspective street, aligned with the
camera view.
4. Limitations and Conclusion
In the course of developing the system we have learned
about some of its limitations. First, the system is as good
a) Camera translation
Original images
Aligned and composite image
Masks and seams between images
Final sequence
b) Camera rotation
Original images
Final sequence
Figure 8: This figure illustrates sequences generated from a large database of street images by inducing two camera motions: a) Camera
translation. Each consecutive image was obtained by inducing camera translation by half the image size as illustrated in figure 2. b) Camera
rotation. Each consecutive image was obtained by inducing camera rotation of 45 degrees as illustrated in figure 2. This produces changes in
the perspective between consecutive images.
as the underlying database is. The larger the database, the
better the results. We also found that there are three typical
failure modes. The first is when a semantically wrong image
is retrieved. This is mitigated by the use of themes but still
remains a problem for future research. Second, compositing two distinct images is always a challenge and at times,
the seam is quite visible. Finally, there are cases in which
the seam runs through important objects in the image which
produces noticeable artifacts.
Nevertheless, the proposed system offers an intuitive 3Dbased navigation approach to browsing large photo collections. This can be used to create photorealistic visual content
for large online virtual 3D worlds like Second Life, computer games or movies. Our system arranges the images into
themes and relies on image content to retrieve the next image. Another novelty of our system, as opposed to existing
image retrieval systems, is the use of transformed image retrieval. We first transform the query image, before performing the query, to simulate various camera motions. Our in-
frastructure also allows us to create infinite panoramas or use
the image taxi to generate tailor-made tours in the virtual 3D
space. These applications can find use in games, movies and
other media creation processes.
References
[1] A. Agarwala, M. Agrawala, M. Cohen, D. Salesin, and R. Szeliski.
Photographing long scenes with multi-viewpoint panoramas. ACM
Trans. Graph., 2006.
[2] A. Agrawal, R. Raskar, and R. Chellappa. What is the range of surface
reconstructions from a gradient field? In Proc. European Conf. Computer Vision, 2006.
[3] S. Avidan and A. Shamir. Seam carving for content-aware image retargeting. ACM Trans. Graph., 2007.
[4] S. Bae, S. Paris, and F. Durand. Two-scale tone management for
photographic look. ACM Transactions on Graphics, 25(3):637–645,
2006.
[5] A. Bosch, A. Zisserman, and X. Munoz. Scene classification using a
hybrid generative/discriminative approach. IEEE Trans. Pattern Analysis and Machine Intelligence, 30(4), 2008.
[6] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction
to Algorithms. The MIT Press, 2nd edition, 2001.
Figure 10: Various panoramas created by our system. The top two panoramas were created automatically from the ‘landscape’ and ‘skyline’
themes respectively. The bottom three panoramas were created interactively from the ‘people’ ‘forest’ and ‘graffiti’ themes respectively.
[7] J. Coughlan and A. L. Yuille. Manhattan world: Orientation and outlier detection by bayesian inference. Neural Computation, 15:1063–
1088, 2003.
[8] J. S. De Bonet. Multiresolution sampling procedure for analysis and
synthesis of texture images. In Proc. ACM SIGGRAPH, 1997.
[9] J. Deutscher, M. Isard, and J. MacCormick. Automatic camera calibration from a single manhattan image. In Proc. European Conf. Computer Vision, pages 175–188, 2002.
[10] A. A. Efros and W. T. Freeman. Image quilting for texture synthesis
and transfer. In Proc. ACM SIGGRAPH, 2001.
[11] W. T. Freeman, T. R. Jones, and E. C. Pasztor. Example-based superresolution. IEEE Computer Graphics & Applications, 22(2):56–65,
2002.
[12] N. Gershenfeld. The nature of mathematical modeling. Cambridge
University, 1998.
[13] J. Hays and A. A. Efros. Scene completion using millions of photographs. ACM Transactions on Graphics, 26(3):4:1–4:7, 2007.
[14] A. Hertzmann, C. E. Jacobs, N. Oliver, B. Curless, and D. H. Salesin.
Image analogies. In Proc. ACM SIGGRAPH, pages 327–340, 2001.
[15] H. D. J. and J. R. Bergen. Pyramid-based texture analysis/synthesis.
In Proc. ACM SIGGRAPH, 1995.
[16] J. Kopf, M. Uyttendaele, O. Deussen, and M. Cohen. Capturing and
viewing gigapixel images. ACM Transactions on Graphics, 26(3),
2007.
[17] V. Kwatra, A. Schödl, I. Essa, G. Turk, and A. Bobick. Graphcut
textures: image and video synthesis using graph cuts. In Proc. ACM
SIGGRAPH, 2003.
[18] J. F. Lalonde, D. Hoiem, A. A. Efros, C. Rother, J. Winn, and A. Criminisi. Photo clip art. ACM Transactions on Graphics, 26(3):3:1–3:10,
2007.
[19] S. Lazebnik, C. Schmid, and J. Ponce. Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories. In
Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2006.
[20] B. D. Lucas and T. Kanade. An iterative image registration technique
with an application to stereo vision. In Proc. of the 7th Int. Joint Conf.
on Artificial Intelligence, pages 674–679, 1981.
[21] A. Oliva and A. Torralba. Modeling the shape of the scene: a holistic representation of the spatial envelope. Int. Journal of Computer
Vision, 42(3):145–175, 2001.
[22] S. Peleg, B. Rousso, A. Rav-Acha, and A. Zomet. Mosaicing on
adaptive manifolds. IEEE Trans. Pattern Analysis and Machine Intelligence, 22(10):1144–1154, 2000.
[23] P. Perez, M. Gangnet, and A. Blake. Poisson image editing. ACM
Trans. Graph., 22(3):313–318, 2003.
[24] P. Rademacher and G. Bishop. Multiple-center-of-projection images.
ACM Trans. Graph., 1998.
[25] C. Rother, L. Bordeaux, Y. Hamadi, and A. Blake. Autocollage. In
Proc. ACM SIGGRAPH, 2006.
[26] Y. Rubner, C. Tomasi, and L. J. Guibas. Adaptive color-image embeddings for database navigation. In ‘Asian Conference on Computer
Vision, pages 104–111, 1998.
[27] N. Snavely, S. M. Seitz, and R. Szeliski. Photo tourism: Exploring
photo collections in 3d. In Proc. ACM SIGGRAPH, 2006.
[28] R. Szeliski. Image alignment and stitching: A tutorial. Foundations
and Trends in Computer Graphics and Computer Vision, 2006.
[29] A. Torralba, R. Fergus, and W. T. Freeman. Tiny images. Technical
Report MIT-CSAIL-TR-2007-024, MIT, 2007.
[30] A. Torralba and P. Sinha. Statistical context priming for object detection. In Proc. IEEE Int. Conf. Computer Vision, pages 763–770,
2001.
[31] Y. Weiss and W. Freeman. What makes a good model of natural images? In Proc. IEEE Conf. Computer Vision and Pattern Recognition,
2007.
[32] D. N. Wood, A. Finkelstein, J. F. Hughes, C. E. Thayer, and D. Salesin.
Multiperspective panoramas for cel animation. ACM Trans. Graph.,
1997.